This project was created for 611 class, fall 2024. The data set I am using is a eye tracking data from 45 healthy adults. During the data collection participants read 6 short texts twice from a given perspective (of a real estate buyer or a baby sitter), then switched their perspective and read the texts again. The data set consists of 39 variables and 76517 observations. My goal is to see the patterns in the data which will help us understand the mechanisms of text processing better. Namely I am interested to look into eye tracking processing data which might predict attention allocation during text rereading.
setwd("~/work")
library(tidyverse);## -- Attaching core tidyverse packages ------------------------ tidyverse 2.0.0 --
## v dplyr 1.1.4 v readr 2.1.5
## v forcats 1.0.0 v stringr 1.5.1
## v ggplot2 3.5.1 v tibble 3.2.1
## v lubridate 1.9.3 v tidyr 1.3.1
## v purrr 1.0.2
## -- Conflicts ------------------------------------------ tidyverse_conflicts() --
## x dplyr::filter() masks stats::filter()
## x dplyr::lag() masks stats::lag()
## i Use the conflicted package (<http://conflicted.r-lib.org/>) to force all conflicts to become errorslibrary(ggplot2);
cleaned_data <- read.csv ("./cleaned_data2.csv");SentenceRelevance_labels <- c("0 Neutral", "1 Buyer", "2 Sitter")
SubjectPerspective_labels <- c("Started with Sitter", "Started with Buyer")
CongruencyLevels_labels <- c("Non-congruent", "Neutral", "Congruent")
cleaned_data$CongruencyLevels <- factor(cleaned_data$CongruencyLevels)The data set consists of eye-tracking measures such as fixation durations, gaze durations, and regression durations. These variables are often correlated, and PCA allows us to condense this information into a smaller number of uncorrelated principal components. I integrated the PCA scores with relevant contextual variables (e.g., “Reading”) for further analysis. PC1: Dominated by variables such as GZD (0.609), TTime (0.466), and FFdur (0.447). These variables likely represent overall gaze duration and total time spent, suggesting PC1 reflects overall reading engagement or attention. PC2: Significant contributions from F2dur (0.548) and FPNfix (-0.542). This suggests PC2 may capture patterns in rereading behavior, distinguishing between second-pass fixations and first-pass fixation transitions. PC3: Primarily driven by IA_First_Fixation_Time (0.699) and Skip (-0.663). This component may represent initial reading strategies, with higher values indicating longer initial fixations and lower values suggesting skipping tendencies. I will focus on PCs 1–3: These explain nearly 60% of the variance and are interpretable in terms of high-level reading strategies like attention allocation, rereading behavior, and initial fixation tendencies.
numeric_cols <- dplyr::select(cleaned_data, IA_First_Fixation_Time, FFdur, F2dur, FPNfix, GZD, regrdur, TTime, Skip)
numeric_cols_scaled <- scale(numeric_cols)
numeric_cols_scaled_clean <- numeric_cols_scaled[complete.cases(numeric_cols_scaled) &
!apply(numeric_cols_scaled, 1, function(x) any(is.infinite(x))), ]
pca_result <- prcomp(numeric_cols_scaled_clean, center = TRUE, scale. = TRUE)
summary(pca_result)## Importance of components:
## PC1 PC2 PC3 PC4 PC5 PC6 PC7
## Standard deviation 1.5188 1.1686 1.0537 0.9995 0.9398 0.9136 0.67017
## Proportion of Variance 0.2883 0.1707 0.1388 0.1249 0.1104 0.1043 0.05614
## Cumulative Proportion 0.2883 0.4590 0.5978 0.7227 0.8331 0.9374 0.99355
## PC8
## Standard deviation 0.22710
## Proportion of Variance 0.00645
## Cumulative Proportion 1.00000pca_result$rotation## PC1 PC2 PC3 PC4
## IA_First_Fixation_Time 0.01128081 -0.1856810 0.69930259 0.27025185
## FFdur 0.44679376 0.0570400 0.02596974 0.29994887
## F2dur 0.25710100 0.5480737 0.16948165 0.06398679
## FPNfix 0.37824933 -0.5422693 -0.16473380 -0.30448658
## GZD 0.60907500 -0.2549171 -0.08242947 -0.04180739
## regrdur 0.05041756 0.3689164 -0.05419034 -0.72068647
## TTime 0.46553590 0.3889549 0.07089244 0.07904674
## Skip -0.02892479 0.1243391 -0.66308616 0.46130575
## PC5 PC6 PC7 PC8
## IA_First_Fixation_Time -0.051389872 0.63291191 -0.01076196 0.0009454112
## FFdur -0.647853415 -0.19537914 0.27046393 -0.4188558581
## F2dur 0.566801895 -0.00347724 0.50948824 -0.1408588771
## FPNfix 0.343767818 0.17414914 -0.04526499 -0.5408039806
## GZD -0.008680241 0.02503446 0.20717001 0.7154585663
## regrdur -0.361883241 0.44709835 0.09022667 -0.0059937050
## TTime 0.082428880 -0.01286963 -0.78324800 -0.0166934411
## Skip 0.019364539 0.57464790 0.02384443 -0.0072905416pca_scores <- pca_result$x
cleaned_data_aligned <- cleaned_data[complete.cases(numeric_cols_scaled) &
!apply(numeric_cols_scaled, 1, function(x) any(is.infinite(x))), ]
pca_plot_data <- cbind(pca_scores, Reading = cleaned_data_aligned$Reading)
pca_plot_data <- as.data.frame(pca_plot_data)A biplot combines the scores of the observations projected onto the principal components and the loadings of the original variables on these components.
plot0 <- biplot(pca_result, scale = 0)ggsave("biplot.png", plot=plot0)## Saving 7 x 5 in imageThese plots visualize the relationship between the first three principal components (PC1 and PC2) and groups the observations (participants) by the variable Reading. The goal is to identify patterns or clusters in the data based on these two dimensions.
plot <- ggplot(pca_plot_data, aes(x = PC1, y = PC2)) +
geom_point(aes(color = Reading), size = 3) +
labs(
title = "PCA: PC1 vs PC2",
x = "Principal Component 1",
y = "Principal Component 2",
color = "Reading"
) +
theme_minimal()
ggsave("figures/PC1vsPC2.png", plot=plot)## Saving 7 x 5 in imageplot2 <- ggplot(pca_plot_data, aes(x = PC1, y = PC3)) +
geom_point(aes(color = Reading), size = 3) +
labs(
title = "PCA: PC1 vs PC3",
x = "Principal Component 1",
y = "Principal Component 3",
color = "Reading"
) +
theme_minimal()
ggsave("figures/PC1vsPC3.png", plot=plot2)## Saving 7 x 5 in imageplot3 <- ggplot(pca_plot_data, aes(x = PC2, y = PC3)) +
geom_point(aes(color = Reading), size = 3) +
labs(
title = "PCA: PC2 vs PC3",
x = "Principal Component 2",
y = "Principal Component 3",
color = "Reading"
) +
theme_minimal()
ggsave("figures/PC2vsPC3.png", plot=plot3)## Saving 7 x 5 in imageAfter reducing the dimensionality of the data using PCA, I am performing clustering to group participants based on their reading behaviors. The goal is to identify distinct patterns or groups that could explain differences in how participants process text during rereading.
The reduced data consists of the first two principal components (PC1 and PC2), which account for 45.90% of the variance in the dataset. These components are used as input for the clustering to simplify the data while retaining most of its meaningful structure.
I chose three clusters as a starting point because the experiment has three main labels or conditions: Neutral, Buyer, and Sitter. By aligning the clustering output with these labels, I can explore whether participants’ reading behaviors naturally align with these experimental conditions or if other patterns emerge.
reduced_data <- as.data.frame(pca_scores[, 1:2])
set.seed(123)
kmeans_result <- kmeans(reduced_data, centers = 3, nstart = 25)
reduced_data$Cluster <- as.factor(kmeans_result$cluster)
kmeans_result$centers## PC1 PC2
## 1 -0.7840434 0.4310042
## 2 0.4285204 -1.0874878
## 3 3.0017653 0.7794242##plot K-means
In this step, I visualized the results of the k-means clustering applied to the principal component scores (PC1 and PC2). The plot color-codes each data point by its assigned cluster and shows their distribution along the first two principal components.
plot4 <- ggplot(reduced_data, aes(x = PC1, y = PC2, color = Cluster)) +
geom_point(size = 3) +
labs(
title = "K-means Clustering on Principal Components",
x = "Principal Component 1",
y = "Principal Component 2",
color = "Cluster"
) +
theme_minimal()
ggsave("figures/K-means.png", plot=plot4)## Saving 7 x 5 in imageIn this analysis, I used principal component analysis (PCA) and k-means clustering to explore eye-tracking data from 45 healthy adults. The goal of the project was to understand patterns in participants’ attention allocation during text rereading, specifically examining how different eye-tracking measures relate to reading behaviors.
I plan to investigate whether these clusters align with experimental conditions, such as the “Reading” condition (i.e., buyer or sitter perspectives), to see if certain reading patterns emerge. I will explore alternative clustering techniques or different numbers of clusters (adjusting the value of k in k-means) to see if better-defined groupings emerge. I will continue by examining how these clusters relate to other variables in the dataset (e.g., congruency levels, sentence relevance) to gain deeper insights into how participants allocate their attention during reading.
This analysis serves as a first step in uncovering patterns in how people read and process text.